Check

Check tells you how closely something follows your brand, and exactly where it doesn't. Paste a link, drop an image or a file, or capture screens from an app, and you get an overall score, a score per subject, and each finding pinned to where it happens, with the brand value it should have used.

It is the same grading your agents get over MCP, so a report means the same thing whether a person or an agent ran it.

Run a check#

1. 1

   Open Check

   Press Check in the sidebar. The page is titled Check against your brand, with one field under it.

2. 2

   Give it something to check

   Paste or drop it into the field that says Paste a link, code or attach a file. The line under the field says what it took the input for, and for visual checks, about how many credits it will cost.

3. 3

   Send

   Press Enter or the arrow. A page check takes about half a minute, longer on a slow site. You can leave the page: the check finishes anyway and waits in the list under the field.

What you can check#

Apps behind a sign-in#

Archetype never signs in to your product. Instead, press Share a tab and choose the tab your app is open in. Then capture each screen you want checked, up to 6 per check. In Chrome and Edge a small floating window carries the capture button over your app, so you can click through it and capture as you go. The capture happens in your own browser, so staging sites, VPN-only apps and localhost work too.

Read the report#

The report leads with a stamp: the overall adherence as a percentage, marked On brand (85 and above), Almost (70 to 84) or Off brand (below 70). A one-line verdict names the subjects that pulled it down, and a Held to chip says which brand it was checked against and how much of it: its colors, its typefaces and, for visual checks, its kit, logo and imagery.

By subject#

Each subject has its own score, in this order where it applies: Page sections, Colour, Typography, Components, Logo, Imagery and Copy. Every score is marked as one of two kinds:

• Counted: colors, type, components and copy are measured. Of the things the check found, it counts how many are the brand's. A value a digit or two off a brand value is a Near-miss and counts three-quarters.
• Judged by eye: page sections, logo and imagery are looked at by Claude, and scored 0 to 100.

On a page, buttons, fields and text styles are measured property by property (height, padding, radius, colors, font, size, weight, line height and more) against the closest variant in your kit and the closest style in your type scale, so even a 1px difference in radius is reported.

Subjects open as an accordion; Expand all opens every one. Findings are numbered on the preview of what you checked, on the right, and clicking a pin opens its subject.

Finding types#

• Not in brand: a value your brand doesn't have.
• Near-miss: almost a brand value, but not quite.
• Hardcoded: in code, a literal that matches a brand token exactly and should reference the token instead.

Claude can only point at fixes your brand actually owns, never an invented value.

Credits and permissions#

Checks of a page, an image or screens include a visual review by Claude and spend AI credits. The estimate shows under the field before you send. Code and copy checks don't show an estimate.

Members without the Designs and checks permission can still run checks, but they can't spend credits, so there is no visual review: the report notes that logo, components and imagery weren't looked at.

History#

Every check the brand has had is listed under the field, newest first, with its score. That includes checks agents ran, marked Agent, and checks run over the API, marked API. Open any row to read its report again or send its link to a teammate. Use the row's menu to delete it.

From an agent#

Agents run the same check with the check_against_brand MCP tool, and the brand's instructions ask them to before they call work done.

• Pass subject with code, CSS, markup or a style object, and get back a worst-first fix list. Each finding says whether it is drift (the brand has a value; use it), a hardcoded literal (reference the token), or a gap (the brand has no value; report it).
• Add kind: "copy" for prose, and each sentence is also checked against the brand's voice documents.
• Pass url instead to check a live page. That takes up to a minute, spends credits, keeps the report on the Check page, and returns a reportUrl to share.
• List anything the agent had to invent in gaps, and it reaches your team.

See MCP tools for the full reference.

Next steps#